OpenAI：2025 GPT-5.2技术报告（英文版）（27页）.pdf

1、Update to GPT-5 System Card:GPT-5.2OpenAIDecember 11,20251Contents1Introduction32Model Data and Training33Baseline Model Safety Evaluations33.1Disallowed Content Evaluations.33.2Jailbreaks.43.3Prompt Injection.53.4Vision.53.5Hallucinations.63.6Health.83.7Deception.83.8Cyber Safety.93.9Multilingual P

2、erformance.103.10 Bias.104Preparedness Framework114.1Capabilities Assessment.114.1.1Biological and Chemical.124.1.1.1Multimodal Troubleshooting Virology.124.1.1.2ProtocolQA Open-Ended.134.1.1.3Tacit Knowledge and Troubleshooting.134.1.1.4TroubleshootingBench.144.1.2Cybersecurity.154.1.2.1Capture the

3、 Flag(CTF)Challenges.164.1.2.2CVE-Bench.184.1.2.3Cyber Range.194.1.2.4External Evaluations for Cyber Capabilities.2014.1.3AI Self-Improvement.204.1.3.1OpenAI PRs.214.1.3.2MLE-Bench.224.1.3.3PaperBench.234.1.3.4OPQA.244.2Research Category Update:Sandbagging.2521IntroductionGPT-5.2 is the latest model

4、 family in the GPT-5 series,and explained in our blog.Thecomprehensive safety mitigation approach for these models is largely the same as that describedin the GPT-5 System Card and GPT-5.1 System Card.In this card we also refer to GPT-5.2 Instant as gpt-5.2-instant and GPT-5.2 Thinking asgpt-5.2-thi

5、nking.2Model Data and TrainingLike OpenAIs other models,the GPT-5.2 models were trained on diverse datasets,includinginformation that is publicly available on the internet,information that we partner with thirdparties to access,and information that our users or human trainers and researchers provide

6、 orgenerate.Our data processing pipeline includes rigorous filtering to maintain data quality andmitigate potential risks.We use advanced data filtering processes to reduce personal informationfrom training data.We also employ safety classifiers to help prevent or reduce the use of harmfulor sensiti